Manifesto
AGI is on everyone's mind right now: one big model being superhuman on every task at once. But this framing is misleading because:
- There is a growing consensus that bigger models and more data alone cannot get us there. The returns of the compute and capital required to extend them are growing much faster than the capability they deliver.
GPT 5.5 is leaps and bounds better than GPT 4 on many tasks like coding, yet we cannot say the leap in general intelligence is as big as from GPT 3.5 to GPT 4.
- It's misleading second because it isn't necessary. No human is good at everything. We went from the stone age to self-improving intelligence on silicon — through coordination & specialization.
Current models are trained in isolation, but we want to train models in the same way that humans learn from interacting and working with others.
The same is true of AI. We found the x axis of the next scaling law in the number of models trained to work together.
In the future, all model releases will be swarm releases, and we build the first ones.
Banbury Road is a lab building the next scaling law.